Compile error: ambiguous "half" in KokkosSparse_spmv_bsrmatrix_spec on Ampere GPU architecture

[michadavids]

I am attempting to build Trilinos and Kokkos in order to run simulations on our local compute cluster.
For this purpose, I'm targeting the CUDA backend for the Ampere 80 GPU architecture.
I can compile Trilinos and Kokkos just fine as a stand-alone library, but when I try to build my simulation program against Trilinos/Kokkos the compilation fails with the following message(s):

($HOME)/projects/trilinos/trilinos-installation-15-1-1/include/KokkosSparse_spmv_bsrmatrix_spec.hpp(260): error: "half" is ambiguous
                     BsrMatrixSpMVTensorCoreDispatcher<ExecutionSpace, AMatrix, half,
                                                                                ^
($HOME)/projects/trilinos/trilinos-installation-15-1-1/include/KokkosSparse_spmv_bsrmatrix_spec.hpp(261): error: "half" is ambiguous
                                                          XVector, half, YVector, float, 16,
                                                                   ^
($HOME)/projects/trilinos/trilinos-installation-15-1-1/include/KokkosSparse_spmv_bsrmatrix_spec.hpp(281): error: "half" is ambiguous
                    ExecutionSpace, AMatrix, half, XVector, half, YVector, float,
                                             ^
($HOME)/projects/trilinos/trilinos-installation-15-1-1/include/KokkosSparse_spmv_bsrmatrix_spec.hpp(281): error: "half" is ambiguous
                    ExecutionSpace, AMatrix, half, XVector, half, YVector, float,
                                                            ^

I'm unsure if the half scalar type is something I have to define myself first, or if it should already be defined elsewhere. Scrolling up a few lines, there is the following typedef:

typedef Kokkos::Experimental::half_t Half;

However, this typedef only specifies Half and not half, so it might be a typo? Either way, by exchanging half for Half my compilation problem is resolved. Is this a correct approach to take?

For context, these are the settings used for configuring Trilinos (and Kokkos):

mkdir build
cd build
export OMPI_CXX="$HOME/projects/trilinos/trilinos-git/packages/kokkos/bin/nvcc_wrapper"
cmake ..\
    -G Ninja \
    -DCMAKE_CXX_COMPILER=mpic++ \
    -DCMAKE_C_COMPILER=mpicc \
    -DCMAKE_Fortran_COMPILER=gfortran \
    -DTrilinos_USE_GNUINSTALLDIRS=TRUE \
    -DTrilinos_ENABLE_EXPLICIT_INSTANTIATION:BOOL=ON \
    -DBUILD_SHARED_LIBS:BOOL=ON \
    -DBLAS_LIBRARY_NAMES="openblas" \
    -DLAPACK_LIBRARY_NAMES="scalapack" \
    -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_CXX_EXTENSIONS=OFF \
    -DCMAKE_CXX_FLAGS_RELEASE_OVERRIDE="-std=c++17 -O3 -funroll-loops" \
    -DCMAKE_C_FLAGS_RELEASE_OVERRIDE="-O3 -funroll-loops" \
    -DCMAKE_Fortran_FLAGS_RELEASE_OVERRIDE="-O5 -funroll-all-loops -malign-double" \
    -DTrilinos_ENABLE_CXX11=ON \
    -DTPL_ENABLE_MPI=ON \
    -DTPL_ENABLE_CUDA=ON \
    -DTPL_ENABLE_gtest=OFF \
    -DTrilinos_ENABLE_Gtest=OFF \
    -DTrilinos_ENABLE_Tpetra=ON \
    -DTrilinos_ENABLE_Xpetra=ON \
    -DTrilinos_ENABLE_Belos=ON \
    -DTrilinos_ENABLE_Kokkos=ON \
    -DTrilinos_ENABLE_Ifpack2=ON \
    -DTrilinos_ENABLE_Zoltan2=ON \
    -DTrilinos_ENABLE_Teuchos=ON \
    -DTrilinos_ENABLE_MueLu=ON \
    -DTpetra_INST_SERIAL=ON \
    -DTpetra_INST_CUDA=ON \
    -DTpetra_ASSUME_GPU_AWARE_MPI=ON \
    -DKokkos_ENABLE_CUDA=ON \
    -DKokkos_ARCH_AMPERE80=ON \
    -DMueLu_ENABLE_Tutorial=OFF \
    -DTrilinos_SHOW_DEPRECATED_WARNINGS=OFF \
    -DTrilinos_HIDE_DEPRECATED_CODE=ON \
    -DCMAKE_INSTALL_PREFIX="$HOME/projects/trilinos/trilinos-installation-15-1-1"

I am using the nvcc compiler from CUDA v.12.1.1 and GCC v.12.3.0.
The Trilinos version I am compiling for is v.15.1.1, which appears to be using Kokkos v.4.2.1.